my-QA-model

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

my-QA-model

This model is a fine-tuned version of distilbert-base-uncased on an SQuAD v1.1 dataset.

Model description

This is a transformer-based extractive Question Answering (QA) model fine-tuned on the Stanford Question Answering Dataset (SQuAD v1.1).
It takes a context paragraph and a natural language question as input and returns the most probable span in the text that answers the question.

• Architecture: DistilBERT
• Dataset: SQuAD v1.1 (~100k question-answer pairs)
• Task Type: Extractive Question Answering
• Training Objective: Predict start and end token positions of the answer span
• Evaluation Metrics: Exact Match (EM) and F1 Score

---

Intended uses & limitations

This model is designed for extractive question answering where the answer exists within a provided context.
It can be applied in reading comprehension tasks, chatbots, document search, automated quiz generation, educational tools, and research on transformer-based QA systems.

However, the model has limitations:

• It can only answer questions if the answer is present in the given text.
• It struggles with multi-hop reasoning, abstract inference, and answers requiring outside knowledge.
• Ambiguous or vague questions may result in incorrect spans.
• Performance may degrade on domains that differ significantly from Wikipedia (SQuAD’s source).
• It may reflect biases in the training data.

Training and evaluation data

The model was fine-tuned on the Stanford Question Answering Dataset (SQuAD v1.1), a large-scale reading comprehension dataset consisting of over 100,000 question–answer pairs on Wikipedia articles.

• Training set: ~87,599 examples
• Validation set: ~10,570 examples
• Each example contains a context paragraph, a question, and the corresponding answer span within the paragraph.

Evaluation was performed on the SQuAD v1.1 validation set using Exact Match (EM) and F1 score metrics.

Training procedure

1. Base Model: A pre-trained transformer model Distibert-base-uncased from Hugging Face.
2. Tokenization: Used the model's corresponding tokenizer with:
   • max_length=384
   • truncation='only_second'
   • stride=128 for sliding window over long contexts
3. Optimization:
   • Optimizer: AdamW
   • Learning rate: 3e-5
   • Weight decay: 0.01
   • Batch size: 16–32 (depending on GPU memory)
   • Epochs: 2–3
4. Loss Function: Cross-entropy loss over start and end token positions.
5. Evaluation: Computed Exact Match (EM) and F1 score after each epoch.
6. Checkpointing: Best model saved based on highest F1 score on validation set.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

The model achieved the following results on the SQuAD v1.1 validation set:

These results are comparable to other transformer-based models fine-tuned on SQuAD , demonstrating strong extractive question answering capabilities.

Framework versions

• Transformers 4.55.0
• Pytorch 2.6.0+cu124
• Datasets 4.0.0
• Tokenizers 0.21.4